DE102022212828A1 - Method and system for automatically generating a function - Google Patents

Abstract

The invention presented relates to a computer-implemented method (100) for automatically generating a function. The method (100) comprises:
- determining (101) a functional safety characteristic value of the function to be generated,
- determining (103) training data from an external source (303) and source information associated with the training data,
- determining (105) a data security index of the training data based on the source information,
- comparing (107) the functional safety index with the data safety index,
- training (109) a machine learner of the function using the training data in the event that the data security index is greater than or equal to the functional security index,
- outputting (111) a directory comprising the source information of respective training data,
- generating (113) the function using the machine learner,
- Setting (115) of a machine (305) by the function.

Description

The presented invention relates to a method and a system for automatically generating a function as well as a corresponding program product according to the appended claims.

So-called artificial intelligence or machine learning is increasingly being used in various fields. Large amounts of data are required to train a machine learner. The training data has a major influence on the structure of a particular machine learner.

Accordingly, a machine learner can be impaired by unsuitable training data, i.e. training data of poor quality. In the worst case, poor-quality data leads to incorrect results.

In contrast to closed systems, in which internal data, i.e. data obtained from sensors in the system, is used to train a machine learner, training data from external sources, such as the Internet, is not subject to complete control by a particular system, so that training data from external sources can be manipulated or incorrect.

Known systems cannot evaluate the quality of training data. Therefore, there is always the risk that machine learning will lead to formally logical but practically incorrect results.

Several testing procedures can be used to validate training data, such as proofreading or peer review, the four-eyes principle (mutual control) or majority decisions in the event of differing results.

Within the scope of the invention presented, a method and a system for automatically generating a function are presented. Further features and details of the invention emerge from the respective subclaims, the description and the drawings. Features and details that are described in connection with the method according to the invention naturally also apply in connection with the system according to the invention and vice versa, so that with regard to the disclosure of the individual aspects of the invention, reference is always made to each other.

Against this background, it is an object of the invention presented to provide a possibility for validating training data for training a machine learner to provide a function for controlling a machine. In particular, the invention presented serves to control an adaptation of a machine learner by training with training data.

Thus, according to a first aspect of the invention presented, a computer-implemented method for automatically generating a function is presented. The presented method comprises determining a functional safety characteristic value of the function to be generated, determining training data and source information associated with the training data from an external source, determining a data safety characteristic value of the training data based on the source information, comparing the functional safety characteristic value with the data safety characteristic value, training a machine learner of the function using the training data in the event that the data safety characteristic value is greater than or equal to the functional safety characteristic value, generating the function using the machine learner, outputting a directory that includes the source information of respective training data and setting a machine through the function.

In the context of the invention presented, an external source is understood to be a source outside of a particular system, i.e. a source that is not subject to the direct control of the computer executing the method presented. An external source can be, for example, a computer network, in particular the Internet.

In the context of the invention presented, source information is understood to mean information about an origin or a generation process or verification process of the respective data. For example, source information can specify a manufacturer or owner of the respective data and/or a verification process underlying the data, such as a peer review or proofreading.

The presented method is based on using a machine learner to generate a function. This means that the function uses a value or decision provided by the machine learner to generate a control signal to control a machine.

In order to control a change in the machine learner and, as a result, the control behavior of the function provided according to the invention, the invention provides that training data for training the machine learner are verified so that, in particular, for example, incorrect or manipulated training data be excluded from influencing the machine learner or the corresponding function.

To verify the respective training data, it is intended that only training data for which source information is available are used.

The respective training data is assessed or verified based on the source information by determining a data security index based on the source information. For example, more or stronger verification processes lead to a better or higher data security index. To assess a verification process, an assignment scheme can be used, for example, which assigns a scale value, such as a numerical value, to the respective verification processes.

As soon as a data security value of the respective training data is known, it can be compared with a functional security value that is specified, for example, as a requirement by a rule or a user. The comparison can be used to determine whether the quality or validity of the training data, expressed by the data security value of a quality requirement, corresponds to that expressed by the functional security value.

In the event that the data security index is at least equal to or better or greater than the functional security index, the training data underlying the data security index can be used to train the machine learner so that the machine learner undergoes a controlled change and influence from, for example, training data manipulated by unauthorized users is excluded.

Furthermore, the invention provides that a directory is output that includes the source information of the respective training data. Using such a directory, a respective machine learner or a corresponding function can also be verified after a training process, so that a user can use the directory to gain confidence in the machine learner or the function and can also integrate it into safety-relevant areas of a system.

It may be provided that the directory includes the data security index of respective training data.

Specifying the data security index in the directory enables a user to make a clear and, if necessary, quantitative assessment of the quality or validity of the training data underlying a particular machine learner.

It can further be provided that the data security index is determined by means of a predetermined assignment scheme, whereby the assignment scheme assigns a respective data security index to respective source information types on the basis of security criteria.

A predefined assignment scheme specifies criteria for assigning a respective data security indicator, such as a level on an ordinal scale, in particular a grade according to a grading scheme, to respective training data or their source information. For example, stronger verification processes or more verification processes lead to a better data security indicator.

It can be provided that the assignment scheme comprises an ascending series of data security indicators and security criteria assigned to respective data security indicators, wherein respective security criteria assigned to a higher-level data security indicator comprise respective security criteria assigned to a lower-level data security indicator.

By formulating security criteria in ascending or mutually referring order, it is ensured that a higher or better data security index includes or quantifies all security criteria of a lower or worse data security index, so that a higher or better data security index can always be used in preference to a lower or worse data security index.

It may be intended that the machine is a vehicle and the function is part of a program classified according to an Automotive Safety Integrity Level (ASIL).

Especially in the area of vehicle development, control over a change or behavior of machine learners is important in order to be able to control a vehicle reliably and predictably in all situations.

It may also be provided that each data security characteristic of the allocation scheme is assigned to an ASIL characteristic.

By assigning respective data security parameters to respective ASIL parameters, the suitability of a respective machine learner or a corresponding function for an application in an area of a software architecture classified according to ASIL can be determined quickly, objectively and reliably.

It may also be provided that the functional safety value is determined based on an ASIL value of the program.

By assigning an ASIL value specified for a function, for example, to a functional safety value, the suitability of a respective machine learner or a corresponding function for an application of the function can be determined quickly, objectively and reliably.

It may further be provided that the source information includes information about a verification process for verifying the training data and the security criteria specify requirements for the verification process.

By using information about a verification process to verify the training data, such as peer review, proofreading or a four-eyes principle, it can be ensured that the training data cannot be manipulated by a single user.

According to a second aspect, the presented invention relates to a system for automatically generating a function, wherein the system comprises a computing unit configured to carry out a possible embodiment of the presented method.

In the context of the invention presented, a computing unit is understood to mean a computer, a processor, a control unit or any other programmable circuit.

According to a third aspect, the presented invention relates to a program product with program code means that configure a computing unit to execute a possible embodiment of the presented method when the program product is executed on the computing unit.

• 1 eine mögliche Ausgestaltung des vorgestellten Verfahrens,
• 2 eine mögliche Ausgestaltung eines Zuordnungsschemas zur Verwendung in dem Verfahren gemäß 1,
• 3 eine mögliche Ausgestaltung des vorgestellten Systems.

Further advantages, features and details of the invention emerge from the following description, in which embodiments of the invention are described in detail with reference to the drawings. The features mentioned in the claims and in the description can each be essential to the invention individually or in any combination. They show:

• 1 a possible design of the presented procedure,
• 2 a possible design of an allocation scheme for use in the procedure according to 1 ,
• 3 a possible design of the presented system.

In 1 A computer-implemented method 100 for automatically generating a function is shown.

The method 100 comprises a first determination step 101 in which a functional safety characteristic value of the function to be generated is determined, for example by determining an ASIL characteristic value of the function.

Furthermore, the method 100 comprises a second determination step 103 in which training data are determined from an external source and source information associated with the training data, for example by downloading training data with a file containing source information from the Internet.

Furthermore, the method 100 comprises a determination step 105 in which a data security characteristic of the training data is determined on the basis of the source information by assigning the source information to a respective data security characteristic according to an assignment scheme.

Furthermore, the method 100 comprises a comparison step 107 in which the functional safety characteristic value is compared with the data safety characteristic value, for example by calculating a difference between the functional safety characteristic value and the data safety characteristic value.

Furthermore, the method 100 comprises a training step 109 in which a machine learner of the function is trained using the training data in the event that the data security index is greater than or equal to the functional security index.

Accordingly, the machine learner is excluded from training using the training data in the event that the data reliability index is smaller than the functional reliability index.

Furthermore, the method 100 comprises an output step 111 in which a directory containing the source information of respective training data is output on an output unit, such as a memory or a display.

Furthermore, the method 100 comprises a generation step 113 in which the function is generated using the machine learner, for example by calling a classification result provided by the machine learner through the function.

Furthermore, the method 100 comprises a setting step 115 in which a machine is set by the function, for example by the function transmitting a control signal to the machine for controlling the machine.

In 2 an assignment scheme 200 is shown. The assignment scheme 200 comprises rows 1 - 6 and columns A - D. Accordingly, the assignment scheme assigns safety criteria specified in column D to respective source information in column C, a corresponding ASIL characteristic value or functional safety characteristic value in column B and a corresponding data safety characteristic value in column A.

Fields A1 - A6 contain consecutive numbers from 1 to 6.

Field B1 states "risk insignificant" and field B2 states "ASIL QM". Fields B3 - B6 list ASIL values A - D in sequence.

Field C1 states "All data from the Internet". Field C2 states "Information on sources". Field C3 states "Information on sources and use of sources of trust level 1-7, verification of the source as a person". Field C4 states "Information on sources and use of sources of trust level 2 --> verification of the source with their intentions and training". Field C5 states "Information on sources and use of sources of trust level 3 --> verification of the source's statement". Field C6 states "Information on sources and use of sources of trust level 4 --> verification of the verified statement of the source".

There is no information in fields D1 and D2. Field D3 states “Trust level 1: traceability to the person”. Field D4 states “Trust level 2 in addition to trust level 1: person has no criminal record and has not posted any conspicuous statements on social media; person is trained or certified as a data labeler”. Field D5 states: “Trust level 3 in addition to trust level 2: data was independently verified (4 or multiple eyes principle); proofreading or peer review; majority decisions in the event of differing results”. Field D6 states: “Trust level 4 in addition to trust level 3: an internal verification process was carried out to check the data (checking data quality, testing, approval)”.

In 3 a system 300 is shown. The system 300 comprises a computing unit 301 which is configured to determine training data from an external source 303, such as the Internet, and to use the training data and corresponding source information to carry out the method 100 according to 1 to control a machine in the form of a vehicle 305.

List of reference symbols

100

Proceedings

101

first investigative step

103

second investigation step

105

Determination step

107

Adjustment step

109

Training step

111

Output step

113

Generation step

115

Setting step

200

Allocation scheme

300

system

301

Computing unit

303

external source

305

vehicle

Claims (10)

Computer-implemented method (100) for automatically generating a function, the method (100) comprising: - determining (101) a functional safety characteristic value of the function to be generated, - determining (103) training data from an external source (303) and source information associated with the training data, - determining (105) a data safety characteristic value of the training data based on the source information, - comparing (107) the functional safety characteristic value with the data safety characteristic value, - training (109) a machine learner of the function using the training data in the event that the data safety characteristic value is greater than or equal to the functional safety characteristic value, - outputting (111) a directory which includes the source information of respective training data, - generating (113) the function using the machine learner, - Setting (115) of a machine (305) by the function. Procedure (100) according to Claim 1 , characterized in that the directory includes the data security index of respective training data. Procedure (100) according to Claim 1 or 2 , characterized in that the data security characteristic value is determined by means of a predetermined assignment scheme (200), wherein the assignment scheme (200) assigns a respective data security characteristic value to respective source information types on the basis of security criteria. Procedure (100) according to Claim 3 , characterized in that the assignment scheme (200) comprises an ascending series of data security characteristics and security criteria assigned to respective data security characteristics, wherein respective security criteria assigned to a higher-level data security characteristic comprise respective security criteria assigned to a lower-level data security characteristic. Method (100) according to one of the preceding claims, characterized in that the machine is a vehicle (305) and the function is part of a program classified according to an Automotive Safety Integrity Level (ASIL). Procedure (100) according to Claim 5 , characterized in that each data security characteristic of the assignment scheme is assigned to an ASIL characteristic. Procedure (100) according to Claim 5 , characterized in that the functional safety value is determined based on an ASIL value of the program. Procedure (100) according to Claim 3 or 4 , characterized in that the source information comprises information about a verification process for verifying the training data and the security criteria specify requirements for the verification process. System (300) for automatically generating a function, wherein the system (300) comprises a computing unit (301) configured to perform a method (100) according to one of the Claims 1 until 8th to execute. Program product with program code means which configure a computing unit (301) to carry out a method (100) according to one of the Claims 1 until 8th to execute when the program product is executed on the computing unit (301).

Images